fix(llm): cap max_tokens and enable streaming to prevent hangs with post-7/31 deepseek-v4-flash (#1204)

[ariesy]

Summary

Fixes #1204. TradingAgents hangs indefinitely with the post-7/31 stable deepseek-v4-flash (long reasoning chains). Root cause is two-fold:

1. No max_tokens sent. _PASSTHROUGH_KWARGS in llm_clients/openai_client.py does not forward max_tokens. With the new long-reasoning v4-flash, an unbounded generation exceeds the gateway's ~3 min idle timeout and the connection is dropped — the process sits in wait_woken forever. When a bounded default does apply, the reasoning chain can consume the entire budget and content comes back empty.
2. No streaming for OpenAI-compatible backends. The OpenCode Go gateway (and similar proxies) closes connections on long non-streamed responses; streaming keeps the socket active.

Verified against opencode.ai/zen/go/v1 with deepseek-v4-flash:

Prompt	max_tokens	Result
trivial	100	OK (reasoning 26 tokens)
complex	500	empty content (reasoning ate all 500)
complex	2048	OK (reasoning 529 + content 55)
complex	unset	no response in 90s (HTTP 000 timeout)

Changes

• tradingagents/default_config.py: add TRADINGAGENTS_MAX_TOKENS env override + max_tokens: 8192 default.
• tradingagents/graph/trading_graph.py: forward max_tokens from config into LLM client kwargs.
• tradingagents/llm_clients/openai_client.py:
  • add max_tokens to _PASSTHROUGH_KWARGS;
  • enable streaming when base_url points at opencode.ai (keeps the socket alive; langchain aggregates chunks back to a single AIMessage);
  • use DeepSeekChatOpenAI for opencode.ai + deepseek models so reasoning_content is captured on the streaming path via _convert_chunk_to_generation_chunk and round-tripped on the next turn.

Verification

• Unit tests: 94 passed (incl. test_deepseek_reasoning, test_capabilities, test_env_overrides, test_openai_compatible_provider, test_llm_max_retries, test_temperature_config).
• End-to-end: full 159566.SZ run (all 4 analysts + research + trader + risk debate) completes; previously hung 3/3 times at the analyst phase.
• max_tokens default (8192) and env override (TRADINGAGENTS_MAX_TOKENS=16384) both verified.

Related

• feat(llm): add OpenCode Go provider #782 (opencode-go provider: streaming required, gateway 3-min idle timeout, reasoning_content delta capture)
• fix(llm): map OpenRouter DeepSeek models to DeepSeek capabilities #1199 (DeepSeek thinking models + tool_choice → very slow responses)